CAKE: a flexible self-supervised framework for enhancing cell visualization, clustering and rare cell identification.

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Single cell sequencing technology offers deep insights into cell heterogeneity.
• High dimensionality and complexity of cell data challenge existing computational clustering methods.
• Current clustering methods lack consistent performance across diverse biological scenarios.

Purpose of the Study:

• To develop a novel, scalable, self-supervised clustering method for analyzing cell heterogeneity.
• To enhance the accuracy, robustness, and biological interpretability of single-cell data clustering.
• To provide improved visualization and identification of major cell types, subgroups, and rare cell populations.

Main Methods:

• Developed CAKE, a self-supervised clustering approach.
• Integrated a contrastive learning model with mixture neighborhood augmentation for cell representation.
• Employed a self-Knowledge Distiller model for refining clustering outcomes.

Main Results:

• CAKE generates condensed and cluster-friendly cell representations.
• Demonstrated superior clustering accuracy and robustness compared to existing methods.
• Successfully identified major cell types, biologically meaningful subgroups, and rare cell types.
• Showcased enhanced visualization capabilities on real single-cell RNA sequencing datasets.

Conclusions:

• CAKE offers a significant advancement in analyzing cell heterogeneity.
• The method provides superior performance in clustering and visualization for single-cell RNA sequencing data.
• CAKE is broadly applicable to cell heterogeneity analysis, particularly for identifying complex cellular structures and rare populations.